The way I have understood it is that SPIRIT is like a project...

The way I have understood it is that SPIRIT is like a project name. It is a project for a device with a fully integrated neural network on-chip with non-volatile resistive memory, so it is going to have multiple iterations and it could use multiple different memories overtime. One of the memories that has been used in demos of the device is Weebit's.

Weebit and Leti have given a demo of the SPIRIT device using Weebit's ReRAM at the Flash Memory Summit 2019 and International Solid- State Circuits Conference (ISSCC) 2020. The demo (MNIST digits recognition) is a hello world type of demonstration, but it is important as it was the first time that it has been done with ReRAM. The system was based on a Spiking Neural Network (SNN) test vehicle for implementing synapses and it showcased the capabilities of Weebit’s technology by performing precise object recognition tasks in an energy-efficient manner. The demo combines CEA-Leti’s SNN algorithms with Weebit’s SiOx ReRAM technology to showcase how circuits can operate similarly to the human brain. It uses bio-inspired architecture to implement synapses in a way that mimics human biological synapse activity, unlike AI circuits that use standard processors and are implemented via software algorithms trying to simulate the synapse function. Using neuromorphic techniques based on ReRAM technology has the potential to emulate the way the brain works, rather than simulate, and makes the computing process for artificial intelligence applications significantly more efficient. This represents a significant long-term opportunity for the use of Weebit’s ReRAM technology, given the significant growth expected in AI applications.

In terms of what the demo does, it recognizes hand written digits. It does this by breaking the writing area into boxes in which it recognizes if there is any writing in there or not. For each digit, it has behind it a model or map that it is building. This model says that if a particular dot is filled then there is a good chance that it is a certain digit or if it is not filled then it is not a certain digit. So for a zero the model has red in the middle as if the middle is filled the chances it is a zero is low. For each digit, it has a separate model.

The demo is not something that anybody wrote or programmed. That is, it learned the models from scratch. In projects like this, there is the training phase and then there is the inference phase. Thousands of people’s handwriting was given to it and the correct digit was told. The ReRAM cell functions like a synapse in your brain and it goes through the same process as a baby does. A baby sees a picture and it says dog and you say no that’s not a dog that’s a cat and it takes that feedback and starts formulating better models of cats and dogs. Over time if they see just a face of a cat, they are able to determine that it is a cat even though they don’t see the full body. Since they have created the model of a cat they can see just the face and then infer the body. The demo project learns by experience just as we learn by experience and as it learns it becomes better

MNIST is a database. The acronym stands for “Modified National Institute of Standards and Technology.” The MNIST database contains handwritten digits (0 through 9), and can provide a baseline for testing image processing systems.

Neuromorphic computing performs computational functions similar to the biological brain to enable AI and machine learning. The suitability of ReRAM for neuromorphic computing is related to the memristor’s ability to modify its state based on the history of voltages applied to it. In other words, it has the temporal and analogue qualities of biological neurons and synapses, i.e. a ReRAM cell not only has an aggregate memory of past inputs but can propagate an output in response to an input based on that aggregate memory, just like a biological brain. As the ReRAM cell structure resembles biological synapses, the technology can potentially be tweaked to serve AI applications in a way that will be much faster and cheaper (less energy consumption) than software-based neural networks. There is a broad future applicability of hardware-based neuromorphic chips in areas such as autonomous driving, advanced driver assistance (ADAS) and edge computing.

https://www.flashmemorysummit.com/Proceedings2019/08-08-Thursday/20190808_AIML-301-1_Regev.pdf